The ABC of Computational Text Analysis

#8 Ethics and the Evolution of NLP

Alex Flückiger

Faculty of Humanities and Social Sciences
University of Lucerne

20 April 2024

Recap last Lecture

  • an abundance of data sources
    • Swissdox, JSTOR, few datasets
  • creating your own dataset
    • convert any data to .txt, incl. OCR
  • processing a batch of files
    • perform tasks in for-loop

Feedback Assignment 2

  • many neat solutions ✅
  • level of sophistication sometimes beyond expectations 🤓
    • highly detailed explanations
    • powerful regex, yet inconsistent

🤔

Do not use ChatGPT naively ⚠️

  1. Copy full assignment into ChatGPT

  2. Get nicely structured, more or less unique output

  3. However, especially for task 1.3, the answers are plain wrong

    • Parsing name in Von BRUNO VANONI, BERN.

🤖 ChatGPT’s Answers (3 attempts, no cherrypicks):

egrep -ho "^[A-Z]+((\s[A-Z]+)*(\s[a-z]+))?,(\s[A-Z]+)+\." newspaper_articles.txt > author_names.txt
egrep -o '^[A-Z][a-z]+ (([A-Z]\.? )?[A-Z][a-z]+ )+' newspaper_articles.txt > authors.txt
egrep -o "^[A-Z][a-z]+ ([A-Z][a-z]+ ?)+," newspaper_articles.txt > author_names.txt

ChatGPT is a tool…

… learn how to use it 👍

  • use ChatGPT as interactive partner
  • don’t trust; try, refine and develop understanding
  • speed up tasks, automation of project not feasible

Example Usage of ChatGPT

“How do I parse two uppercased words at the beginning of a line after the word ‘Von’?”

^Von [A-Z]+ [A-Z]+

“How would it work in Python?”

pattern = r'^Von [A-Z]+ [A-Z]+'
matches = [line for line in lines if re.match(pattern, line)]

Outline

  • ethics is everywhere 🙈🙉🙊
    • … and your responsibility
  • understand the development of modern NLP 🚀
    • … or how to put words into computers

Ethics is more than philosophy.
It is everywhere.

An Example

You are applying for a job at a big company

Does your CV pass the automatic pre-filtering?

🔴 🟢

Your interview is recorded. 😎 🥵
What personal traits are inferred from that?

Face impressions as perceived by a model by (Peterson et al. 2022)

Don’t worry about the future…

… worry about the present.

  • AI is persuasive in everyday’s life
    • assessing risks and performances (credits, job, crimes, terrorism etc.)
  • AI is extremely capable
  • AI is smart within limits only and often poorly evaluated

. . .

💡 What is going on behind the scene?

An (R)evolution of NLP

From Bag of Words to Embeddings

Putting Words into Computers (Smith 2020; Church and Liberman 2021; Manning 2022)

  • from coarse+static to fine+contextual meaning
  • how to measure similarity of words and documents?
  • from counting to learning representations

Bag of Words

  • word as arbitrary, discrete numbers
    • King = 1, Queen = 2, Man = 3, Woman = 4
  • intrinsic meaning
  • how are these words similar?

Vector-representations of words as discrete symbols (Colyer 2016)

Representing a Corpus

Collection of Documents

  1. NLP is great. I love NLP.

  2. I understand NLP.

  3. NLP, NLP, NLP.

Document Term Matrix

NLP I is term
Doc 1 2 1 1 …
Doc 2 1 1 0 …
Doc 3 3 0 0 …
Doc ID … … … term frequency

“I eat a hot ___ for lunch.”

You shall know a word by the company it keeps!

Firth (1957)

Word Embeddings

word2vec (Mikolov et al. 2013)

  • words as continuous vectors
    • accounting for similarity between words
  • semantic similarity
    • King – Man + Woman = Queen
    • France / Paris = Switzerland / Bern

Single continuous vector per word (Colyer 2016)

Words as points in a semantic space (Colyer 2016)

Doing arithmetics with words
(Colyer 2016)

Contextualized Word Embeddings

BERT (Devlin et al. 2019)

  • recontextualize static word embedding
    • different embeddings in different contexts
    • accounting for ambiguity (e.g., bank)
  • acquire linguistic knowledge from language models (LM)
    • LM predict next/missing word
    • pre-trained on loads of data


💥 embeddings are the cornerstone of modern NLP

Large Language Models (LLM)

ChatGPT (OpenAI 2023)

  • scale up attempts of previous models
    • more model parameters (>175B) and train data (>300B words)
  • optimize for conversations
    • instruction-tuning (summarize, translate, reason)
    • Reinforcement Learning from Human Feedback (RLHF)

🤓 There are dozens other models than ChatGPT.

Modern NLP is propelled by Data

Associations in Data


«___ becomes a doctor.»

Learning Patterns from Data

Gender bias of the commonly used language model BERT (Devlin et al. 2019)

Cultural Associations in Training Data

Gender bias of the commonly used language model BERT (Devlin et al. 2019)

Word Embeddings are biased …

… because our data is we are biased. (Bender et al. 2021)

In-class: Exercises I

  1. Open the following website in your browser: https://pair.withgoogle.com/explorables/fill-in-the-blank/
  2. Read the the article and play around with the interactive demo.
  3. What works surprisingly well? What looks flawed by societal bias? Where do you see limits of large language models?

Modern AI = DL

How does Deep Learning look like?

Simplified illustration of a Neural Network. Arrows are weights.

How does Deep Learning work?

Deep Learning works like a huge bureaucracy

  1. start with random prediction
  2. blame units for contributing to wrong predictions
  3. adjust units based on the accounted blame
  4. repeat the cycle

. . .

🤓 train with gradient descent, a series of small steps taken to minimize an error function

Current State of Deep Learning

Extremely powerful but … (Bengio, Lecun, and Hinton 2021)

  • great at learning patterns, yet reasoning in its infancy
  • requires tons of data due to inefficient learning
  • generalizes poorly

Limitations of data-driven Deep Learning


„This sentence contains 37 characters.“
„Dieser Satz enthält 32 Buchstaben.“
 

Doubts about practical implications?

Gender bias in Google Translate

Biased Data and beyond

Raw data is an oxymoron.

Gitelman (2013)

Three Sides of the AI Coin

Explaining vs Solving vs Tracking

  • conduct research to understand
  • automate tedious tasks
  • track people for profit or political reasons

Fair is a Fad

  • companies also engage in fair AI to avoid regulation
  • Fair and good – but to whom? (Kalluri 2020 )
  • lacking democratic legitimacy

Don’t ask if artificial intelligence is good or fair, ask how it shifts power.

Kalluri (2020)

Data represents real life.

Don’t be a fool. Be wise, think twice.

Algorithmic Managment of Labour Force

Text generation may be used to communicate difficult decisions strategically

Questions?

References

Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜.” In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–23. Virtual Event Canada: ACM. https://doi.org/10.1145/3442188.3445922.
Bengio, Yoshua, Yann Lecun, and Geoffrey Hinton. 2021. “Deep Learning for AI.” Communications of the ACM 64 (7): 58–65. https://doi.org/10.1145/3448250.
Church, Kenneth, and Mark Liberman. 2021. “The Future of Computational Linguistics: On Beyond Alchemy.” Frontiers in Artificial Intelligence 4. https://doi.org/10.3389/frai.2021.625341.
Colyer, Adrian. 2016. “The Amazing Power of Word Vectors.” the morning paper. 2016. https://blog.acolyer.org/2016/04/21/the-amazing-power-of-word-vectors/.
Devlin, Jacob, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding.” http://arxiv.org/abs/1810.04805.
Firth, John R. 1957. “A Synopsis of Linguistic Theory, 1930-1955.” In Studies in Linguistic Analysis: Special Volume of the Philological Society, edited by John R. Firth, 1–32. Oxford: Blackwell. http://ci.nii.ac.jp/naid/10020680394/.
Gitelman, Lisa. 2013. Raw Data Is an Oxymoron. Cambridge: MIT.
Kalluri, Pratyusha. 2020. “Don’t Ask If Artificial Intelligence Is Good or Fair, Ask How It Shifts Power.” Nature 583 (7815, 7815): 169–69. https://doi.org/10.1038/d41586-020-02003-2.
Manning, Christopher D. 2022. “Human Language Understanding & Reasoning.” Daedalus 151 (2): 127–38. https://doi.org/10.1162/daed_a_01905.
Mikolov, Tomas, Ilya Sutskever, Kai Chen, Greg S. Corrado, and Jeff Dean. 2013. “Distributed Representations of Words and Phrases and Their Compositionality.” In Advances in Neural Information Processing Systems, 3111–19.
OpenAI. 2023. “GPT-4 Technical Report.” March 27, 2023. http://arxiv.org/abs/2303.08774.
Peterson, Joshua C., Stefan Uddenberg, Thomas L. Griffiths, Alexander Todorov, and Jordan W. Suchow. 2022. “Deep Models of Superficial Face Judgments.” Proceedings of the National Academy of Sciences 119 (17): e2115228119. https://doi.org/10.1073/pnas.2115228119.
Smith, Noah A. 2020. “Contextual Word Representations: Putting Words into Computers.” Communications of the ACM 63 (6): 66–74. https://doi.org/10.1145/3347145.